Graph-Regularized Attentive Convolutional Entanglement for Robust DeepFake Video Detection

The proposed GRACE method leverages feature entanglement with sparse constraints and a graph convolutional network with graph Laplacian smoothing prior regularization to effectively exploit the spatial-temporal correlations in face sequences while suppressing the impact of noise and distortions.

所提出的GRACE方法借助结合稀疏约束的特征纠缠机制，以及搭载图拉普拉斯平滑先验正则化的图卷积网络（Graph Convolutional Network），能够有效挖掘人脸序列中的时空关联特性，同时抑制噪声与畸变的影响。